以DeepSeek‑R1为例，仅靠强化学习训练，模型在AIME数学推理基准上的pass@1从15.6%提升至 77.9%，充分展示了RL在低数据量条件下即可实现大幅能力跃升，迅速成为后训练赛道的新范式。

机器之心发布当 OpenAI 前 CTO Mira Murati 创立的 Thinking Machines Lab (TML) 用 Tinker 创新性的将大模型训练抽象成 forward backward，optimizer step ...

A Polymarket user turned a $33,000 bet into over $400,000, wagering that Venezuela’s President Nicolás Maduro would be ousted before the month’s end.

I really have too many tray icons. You know the ones. They sit on your taskbar, perhaps doing something in the background or, ...

Legal expert Katherine Yon Ebright breaks down the legality of the US strikes in Venezuela and what comes next for the ...

Anthropic's Ralph plugin keeps Claude retrying until specs pass, with a stop hook to pause loops, so you ship cleaner code ...

Australian households can save thousands of dollars annually by auditing bank statements and eliminating wasteful spending on subscription services, gym memberships, and daily purchases.

A Champions Cup so lacking in substance and purpose, it resembles a ghost searching for someone to haunt; a pre-shrunk Lions ...

Weekly roundup exploring how cyber threats, AI misuse, and digital deception are reshaping global security trends.

OpenAI、Anthropic和Google的AI代码助手现在能够在人工监督下连续工作数小时，编写完整应用、运行测试并修复错误。但这些工具并非万能，可能会让软件项目变得复杂。AI代码助手的核心是大语言模型，通过多个LLM协作完成任务。由于存在上下文 ...

The new major version with a new JIT compiler, a revised parallelization API, and a maturing type system paves the way for ...

Anyone who's touched GIMP, Krita, or Inkscape knows Linux is anything but a barren wasteland. The open-source creative world ...